From left to right: Siyi, Zifeng, Tejas, and... me!
This weekend, my team and I took part in Jane Street's Electronic Trading Challenge 2022 and came in at 1st place!
The contest was held in Singapore, at the very luxurious Shangri-La hotel on 17/09/2022 from 10am till 6pm. Importantly, breakfast and lunch were catered (yum)!
Provided by the organizers:
etc is a programming challenge. The goal is to write a program that will make fake money by trading securities on a fake financial exchange. The market participants will include other people’s programs, as well as some programs provided by us (the “marketplace bots”).
For the challenge, Jane Street provided each team with a secured Ubuntu Linux box in the cloud wherefrom we could interact with the production exchange to place buy and sell orders.
They also kindly provided us with a Python3 convenience class Exchange() that exposed general utility functions such as .read_public_feed(), .send_order(order) and .cancel_order(id).
Lastly, there were two deployment environments --production and --test to which we could field our bots.
The event kicked off at 10am sharp with a briefing on what the challenge actually was (it was the first time for all of us).
Provided by the organizers:
The challenge consists of a series of five minute matchups. In each matchup, you and another team compete to be the best market maker, providing liquidity to a third bot called the “marketplace”.
The marketplace will send some orders around the fair value. Some orders will, like yours, be orders intended to provide liquidity. Other orders from the marketplace will be emulating liquidity demanders – people who just wish to buy or sell a certain amount of shares, and don’t care too much about the price they get.
At the end of a matchup, the exchange closes and then reopens shortly afterwards for the next matchup, where you’ll be playing alongside a different team. We reset your cash and positions too.
There will be two distinct competitions, with two distinct prizes.
The first will run (more-or-less) continuously between 11:00 am and 5:00 pm. Scores during each phase of this competition will be weighted as shown below...
At 4:00 pm we will start the second, simultaneous competition: “The Final Hour”. The winner of this competition is the team that has the highest cumulative score in just the last hour of competition.
Then, we were given handouts on various financial instruments like stocks, bonds and ADRs to study (we are programmers, not bankers). There was a lot to read and understand so I busied myself with the docs.
At 11am, the linux boxes were finally opened and, with palpable excitement in the air, we were off to the races!
In the challenge, there were four "regularly stocks" [VALBZ, GS, MS, WFC]. In addition, there was VALE (an ADR of VALBZ) and XLF (an ETF of [GS, MS, WFC]).
Immediately, it was clear to us that the assets divided themselves very naturally into two types: (1) ADRs and (2) ETFs.
Hence, we split ourselves further into two subteams. Zifeng and Si Yi drafted tactics for [VALBZ, VALE] while me and Tejas developed strategies for [GS, MS, WFC, XLF].
Zifeng and Si Yi were the quickest to enter the market; they prototyped their ADR arbitrage strategy and deployed it to production at breakneck speed.
Basically, 1 share of VALE can be converted to/from 1 share of VALBZ (with a conversion fee of $10). Then, if the prices of those two instruments ever differed significantly due to inefficiencies in the market, we could buy the cheaper of the two, convert it into the other and sell for higher!
The results were good too!
We were earning around $300 per round and were consistently ranking between 3rd - 7th!
On our side, the idea was similar.
According to the specs, 10 shares of XLF could be converted to/from a basket of [3 BOND, 2 GS, 3 MS, 2 WFC] (with a conversion fee of $100). Then, if the sum of the underlying assets ever differed significantly from XLF, their aggregate, then we could buy the cheaper of the two, convert into the other and (again) sell for higher.
That said, the details of the implementation were not immediately clear to me so Tejas took on the Mathematics while I collated live data for each round.
The questions of interest included the following: (1) what are the opening and closing prices of each stock for each round and (2) how many trades of each stock took place per round.
It was here that I saw something suspicious.
For a few rounds in a row, one of our bots in the test environment was steadily raking in profits while doing no trading at all!
Someone had bought 2 shares of MS and 12 shares of XLF previously but then forgot about it. That meant that just by holding those two stocks we were contributing towards our P&L via appreciation!
Then, we added a new strategy to production: at the start of each 5 minute round, buy as many shares of MS and XLF as we could and hold them all the way.
The result? Surprising. We were now earning roughly $600 per round and we didn't even know why HAHAHA (is it truly a trading challenge if there wasn't any speculation???)
It was around here, at 2pm, that we advanced to 2nd overall and gained some much needed confidence!
Halfway through the competition, we realized something.
There were roughly 5,000 trades per round. Instead of racking our brains trying to earn the maximum possible profit on a single trade, why not earn a dollar on every one of them? Instead of waiting for the market to be "ripe" for arbitrage (passive/ timing the market), we wanted a cut of every trade ever made (active/ time in the market).
In essence, we wanted to be market makers. For each asset, we would (A) compute its fair price and (B) place a buy order at fair_price - 1 and another sell order at fair_price + 1, i.e., buy low and sell high.
Here, we set the delta to be +/- 1 so as to keep the spread tight and ensure that our orders would always get filled first (and, in fact, every time).
There was just one problem. How do we compute the fair price? The most natural idea would be to take the average of the best buy and sell orders. We did not do that.
We felt that the order book was an aggregation of individual speculation. That is, it was likely not the best method to infer a fair price.
Instead, we took the last transacted price of the asset to be its fair price.
The result? Spectacular! We were now pulling in over $10,000 per round! They say that the biggest winners during the gold rush were the people who sold the shovels. Here? That seemed to check out.
By providing liquidity (always ensuring people could buy or sell their assets) and capturing the bid-ask spread, we pulled away from the competition entirely and were now comfortably in 1st place with two hours left to go.
At last, with our refined implementations all teams started the last leg of the competition.
All cash and positions were reset and the main event commenced.
Since deploying our market making strategy to production, we were earning an average of $10,000 per round. Would that continue to be the case?
Turns out our strategy was sufficient. We lead the event comfortably from the start, though that did nothing to assuage the anxiety that another team might come up from behind to overtake us.
It was only after all twelve rounds of the final event had concluded and we won with a P&L of $8,000,000 that we could breathe a sigh of relief!
This is my first experience with algorithmic trading and it was a great learning experience! If anything, seeing the number of trades per second executed by our trading bot truly opened my eyes to the world and prowess of high-frequency trading.
Seeing our P&L increase steadily at first and then dramatically towards the end was also super exhilarating! That said, trading at a loss was anxiety-inducing in equal measure... the day was a roller coaster-ride to be sure!
For future teams, I will say that the entire challenge will be a time scramble. From understanding how the exchange works, how orders are placed, how feeds are read and how they can be parsed, etc. More or less as first-timers no team will get down a perfect strategy from the get-go.
What is important is being able to adapt and learn the ropes on the fly to (1) develop a mental model of how trading works and (2) having the creativity to try different things! No matter how good you think your strategy is, the only thing that matters is whether it will work! So iterate often, develop quickly and test, test, test!
Try hard, but don't take the event too seriously as well HAHAHA! Go out there and try silly tactics, gamble a bit and have... fun XD
... I hope you [develop a strategy] you’re proud of, and if you’re not, I hope you have the courage to start over again. -- F. Scott Fitzgerald (probably)
Having participated once, I will sadly no longer be eligible for subsequent iterations.
If there was one thing I wish I had done better, that would be to save the logs of every single round into their own .csv files and then fire up pandas to study what happened afterwards. That would have provided greater insight into what other teams might have been trying/ how the stocks were changing hands and how frequently. I might have been able to see how the strategies of my opponents changed over time and maybe counter them somehow hmm...
Overall, I felt the event was extremely well organized and very, very fun! The infrastructure alone must have been a humongous effort behind the scenes. The seamless execution to (1) ensure that the exchange ran smoothly and (2) rate-limited participants appropriately, (3) correctly servicing thousands of orders per minute and (4) broadcasting saids trades and other metadata to public feeds in a timely manner were especially impressive.
Thank you, Jane Street, for the amazing event. It really was memorable beyond words.